Our submitted files for the business track of the 2019 NBA Hackathon. Given a dataset of Instragram posts (time posted, description, followers at time of posting, and type), we predicted engagement with a 3.08% MAPE on the test set. Applicants were ranked by MAPE on a private holdout set. Feature engineering included splititng timestamps into categorical variables then appling sin / cos transformations, extracting number of mentions and hashtags from description, and using SPPMI to embed descriptions. The hyperparameters for the final XGBoost model were found with Bayesian optimization (hyperopt).
jeromecohen / nba Goto Github PK
View Code? Open in Web Editor NEWSuccessful Submission to the 2019 NBA Hackathon